Skip to content
View HenryNdubuaku's full-sized avatar
203 followers · 42 following
  • Cactus Compute
  • London
  • 08:26 (UTC)

Achievements

Achievement: Starstruckx3Achievement: Pair Extraordinairex2Achievement: QuickdrawAchievement: YOLOAchievement: Pull Sharkx2Achievement: Arctic Code Vault Contributor

Achievements

Achievement: Starstruckx3Achievement: Pair Extraordinairex2Achievement: QuickdrawAchievement: YOLOAchievement: Pull Sharkx2Achievement: Arctic Code Vault Contributor

Block or report HenryNdubuaku

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don't include any personal information such as legal names or email addresses. Markdown supported. This note will be visible to only you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
HenryNdubuaku/README.md

Henry Ndubuaku

LinkedIn Twitter Email Spotify

I could train a 1B-A200m model on an iPhone 17 Pro at ~650 tokens/sec. It will take 360 days on 20B tokens of training data and use 156KW of electricity which cost $51.

The phone will fry of course, so I wrote algorithms to run inference on your phone rather. We named it after a plant that survives in resource-constrained environments, Cactus.

cactus can run similar model on your Grandma’s Pixel 6a at 36 tokens/second while only draining 10% battery per hour of continuous inference.

I had a verbal offer from Nvidia, my dream company, but passed on to build this. Cactus now has 3.8k GitHub stars and have completed 6m inference tasks in production.

We raised funding from YCombinator, FCVC, Oxford University, 62 tech CTOs/VP Eng, our fellow YC batchmates, a couple DeepMind engineers and Transpose (Garry Tan's brother).

Core Expertise

Maths Computing AI/ML/RL Distributed Systems GPU

Main Tools

Python C++ PyTorch Jax CUDA Vulkan Neon Cloud

Career Progression

  • 2025-XX: Cactus (YC S25) - Founder & CTO (tiny inference engine for phones and wearables).
  • 2024-25: Deep Render - AI Research Engineer (realtime video models that run on phone GPU/NPU).
  • 2021-24: Wisdm - ML Software Engineer (distributed perception AI for Maxar Defence satelite views).
  • 2019-21: MSc + Open-source activities (JAX/NanoDl, Torch/SuperLazyAutograd, CUDARepo, etc.).
  • 2018-19: Google GADS Scholarship Programme with Andela (pre-MSc), around systems design.
  • 2017-18: National Youth service, posted to software engineering after bootcamp, mostly ARM.
  • 2012-16: Started uni at 15y, covered EECS, data structures, algorithms, maths, physics.

Fun Highlights

  • Wrote Math & CS For ML (with codes).
  • Gave this lecture to a small ML group in Nigeria, on optimising large-scale ML in JAX.
  • Co-host this monthly dinner for AI researchers, engineers and founders in London.
  • Kevin Murphy (DeepMind Principal), Thomas Wolf (HuggingFace Co-foubder), Daniel Holtz (Mid Journey Founder), Steve Messina (IBM CTO) followed back on X.
  • After CUDARepo, Nvidia reached out, I did 7 technical rounds, got a verbal offer, back-and-forth over YOE/pay, then I got YC.
  • Did MSc at QMUL, just to work with Prof Matt Purver (Ex-Stanford Researcher on CALO), did my project/thesis with his team.
  • Did BEng under Prof Onyema Uzoamaka (Rumoured first Nigerian CS grad from MIT), he taught computing archs off-head!

Pinned Loading

  1. cactus-compute/cactus cactus-compute/cactus Public

    Kernels & AI inference engine for mobile devices.

    C++ 3.8k 229

  2. nanodl nanodl Public

    A Jax-based library for building transformers, includes implementations of GPT, Gemma, LlaMa, Mixtral, Whisper, SWin, ViT and more.

    Python 297 12

  3. cuda-tutorials cuda-tutorials Public

    CUDA tutorials for Maths & ML tutorials with examples, covers multi-gpus, fused attention, winograd convolution, reinforcement learning.

    Cuda 203 8

  4. super-lazy-autograd super-lazy-autograd Public

    Hand-derived memory-efficient super lazy PyTorch VJPs for training LLMs on laptop, all using one op (bundled scaled matmuls).

    Python 39

  5. pete pete Public

    Parameter-efficient transformer embeddings replace learned embeddings with hardware-aware polynomial expansions of token IDs.

    Python 6

  6. tango tango Public

    Decentralised ML engine, where tiny edge devices like smart watches, phones, VR headsets, game consoles etc., could contribute.

    Go 3 1